New Framework Maps AI Control by Technical Capability, Not Market Power
An Indian report proposes assigning AI responsibilities based on technical control, arguing India's current regulations incorrectly target companies lacking true oversight. The framework aims to align obligations with actual capabilities.
A new report from Indian think tank The Quantum Hub suggests a functional framework for assigning responsibility for harms caused by artificial intelligence (AI). The report argues that India's current regulatory approach, which relies on broad categories like 'intermediary', misattributes obligations to companies that lack the technical means to control AI outputs.
The report, titled "Functional Taxonomy of Actors in the AI Value Chain," breaks down the AI ecosystem into eight actor classes and five functional attributes. By mapping these, it aims to identify which actor has the technical ability to prevent, mitigate, or respond to a specific risk, thereby ensuring that obligations are placed where actual control exists.
How it Works:
The taxonomy identifies eight actor classes, including physical infrastructure providers, cloud providers, data actors, foundation model developers, orchestration providers, model adaptors, distribution platforms, and end-users. These are mapped against five functional attributes representing forms of control: model layer influence, runtime control, output visibility, distribution control, and end-user proximity. The report emphasizes that a company might occupy multiple positions, each with different control capabilities.
The core principle is function over entity. The report posits that placing an obligation on an actor without the corresponding technical control results in compliance costs without safety benefits. It highlights that control after a model's release varies significantly, especially differentiating between closed API models and open-weight or open-source releases.
Limitations:
The framework specifically addresses generative AI outputs like text, audio, and video, excluding autonomous decision-making systems or real-world actions not directly tied to content generation. Critically, the report acknowledges it does not account for market power or how a single, vertically integrated company might hold control across multiple categories. This concentration of power is a governance fact that the functional taxonomy does not capture.
The report aims to provide a more precise basis for AI regulation in India, but leaves open questions about application to open-source models and situations where control is diffused or difficult to locate, potentially allowing well-advised companies to route around regulations.